Multivariate spatial conditional extremes - Lancaster University

Page created by Nancy Moreno
 
CONTINUE READING
Multivariate spatial conditional extremes - Lancaster University
Multivariate spatial conditional extremes

Rob Shooter, Emma Ross, Agustinus Ribal, Ian Young,
                 Philip Jonathan

                          MetOffice, UK
                      Shell, The Netherlands
           Hasanuddin University, Makassar, Indonesia
               University of Melbourne, Australia
               Shell and Lancaster University, UK.

                    RSS Manchester
(Slides and draft paper at www.lancs.ac.uk/∼jonathan)

Jonathan                     MSCE                       September 2021   1 / 23
Multivariate spatial conditional extremes - Lancaster University
Introduction

Acknowledgement, motivation, related work
Acknowledgement
 ◦ Jon Tawn and Jenny Wadsworth (Lancaster), David Randell (Shell)
Motivation
 ◦ How useful are satellite observations of ocean waves and winds?
 ◦ Could they become the primary data source for decisions soon?
 ◦ What are the spatial characteristics of extremes from satellite
   observations?
Related work
 ◦ Heffernan and Tawn [2004] (CE), Heffernan and Resnick [2007]
 ◦ Shooter et al. [2019] (SCE), Wadsworth and Tawn [2019] (SCE)
 ◦ Shooter et al. [2021b], Shooter et al. [2021a] (SCE applications)
Competitors (= MSPs, hierarchical MSPs and multivariate MSPs)
 ◦ Reich and Shaby [2012], Vettori [2017], Vettori et al. [2019]
 ◦ Genton et al. [2015], Huser and Wadsworth [2020]
       Jonathan                       MSCE            September 2021   2 / 23
Multivariate spatial conditional extremes - Lancaster University
Introduction

Summary of talk : Outline

 ◦ A look at the data
 ◦ Brief overview of methodology, extended to multiple fields
 ◦ Results for joint spatial structure of extreme scatterometer wind
   speed, hindcast wind speed and hindcast significant wave height
   in the North Atlantic
 ◦ Implications for future practical applications

      Jonathan                      MSCE            September 2021   3 / 23
Multivariate spatial conditional extremes - Lancaster University
Introduction

Summary of talk : Methodology in nut-shell
                 ◦ Condition on large value x of first quantity X01 at
                   one location j = 0
                 ◦ Estimate “conditional spatial profiles” for m > 1
                                       p,m
                   quantities { X jk } j=1,k=1 at p > 0 other locations

                                          X jk ∼ Lpl
                                            x>u
                                 X |{ X01 = x} = α x + xβ Z
                               Z ∼ DL(µ , σ 2 , δ ; Σ(λ , ρ, κ ))
                 ◦ MCMC to estimate α , β, µ , σ , δ and ρ, κ , λ
                 ◦ α , β, µ , σ , δ spatially smooth for each quantity
                 ◦ Residual correlation Σ for conditional Gaussian
                   field, powered-exponential decay with distance

      Jonathan                   MSCE                    September 2021   4 / 23
Multivariate spatial conditional extremes - Lancaster University
Data    Basics

Satellite observation
                                         Features
                                            ◦ Altimetry: HS and U10
                                            ◦ Scatterometry: best for U10 and
                                              direction
                                            ◦ > 30 years of observations
                                            ◦ Spatial coverage is by no means
                                              complete: one observation daily
                                              if all well
                                            ◦ Calibration necessary (to buoys
                                              and reanalysis datasets, Ribal
                                              and Young 2020)
                                            ◦ METOP(-A,-B,-C) since 2007
                                         HS : significant wave height (m)
         [Ribal and Young 2019]          U10 : wind speed (ms−1 ) at 10m (calibrated to 10-minute average
                                         wind speed)

      Jonathan                       MSCE                                 September 2021             5 / 23
Multivariate spatial conditional extremes - Lancaster University
Data   Basics

Hindcast data, objectives
Hindcast data
  ◦ Physical simulator calibrated to observations (e.g from buoys)
  ◦ NORA10 hindcast covers North Atlantic off UK (Breivik et al.
     2013)
  ◦ Data available 1957-2018
Initial objective
  ◦ Joint spatial inferences about extremes using all of
      ◦ HS (JASON)
      ◦ directional U10 (METOP)
      ◦ directional HS and directional U10 (NORA10)
  ◦ Not feasible: poor joint spatial coverage of JASON and METOP
Revised objective
Joint spatial inferences about extremes of directional U10 (METOP),
hindcast directional HS and directional U10 (NORA10)
       Jonathan                   MSCE                September 2021   6 / 23
Multivariate spatial conditional extremes - Lancaster University
Data   Basics

JASON and METOP

  [n2yo.com, accessed 06.09.21 at around 1100UK]             [stltracker.github.io, accessed 27.08.2021 at around 1235UK]

 ◦ JASON and METOP similar polar orbits
 ◦ JASON all ascending, METOP all descending over North Atlantic
 ◦ Joint occurrence of JASON and METOP over North Atlantic rare
         Jonathan                                     MSCE                                 September 2021            7 / 23
Data     Preprocessing

Swath wind speeds

     Daily descending METOP swaths. Satellite swath location changes over time. Spatial structure evident

     Jonathan                                      MSCE                                 September 2021      8 / 23
Data   Preprocessing

Registration locations

                                                                   Procedure
                                                                      ◦ 14 longitude-latitude pairs
                                                                      ◦ Satellite observation nearest to
                                                                        each pair used for each swath
                                                                      ◦ Corresponding hindcast data for
                                                                        each pair at time of swath
                                                                      ◦ “Instantaneous” satellite wind
                                                                        vector, hindcast wind vector,
                                                                        hindcast HS and wave direction
                                                                        for 1532 times
                                                                      ◦ Most southerly location for
                                                                        conditioning in MSCE
 Registration locations : square is conditioning location
                                                                      ◦ Note colour scheme
  StlWnd (green), HndWnd (orange), HndWav(blue)

            Jonathan                                           MSCE                   September 2021   9 / 23
Data     Registered data

Scatter plots on physical scale

             Scatter plots of registered data : StlWnd (green), HndWnd (orange), HndWav(blue)

      Jonathan                                   MSCE                                 September 2021   10 / 23
Data     Registered data

Covariate dependence

   Directional and seasonal dependence. “Direction” is that from which fluid flows measured clockwise from North
                                 StlWnd (green), HndWnd (orange), HndWav(blue)

         Jonathan                                     MSCE                                September 2021           11 / 23
Data   Marginal transformation to Laplace scale

Marginal transformation to standard Laplace scale
Procedure
  ◦ Non-stationary piecewise constant directional-seasonal marginal
    extreme value model (github.com/ECSADES/ecsades-matlab)
  ◦ Pre-specified 8 directional bins (“octants”) of equal width centred
    on cardinal and semi-cardinal directions
  ◦ Pre-specified “summer” and “winter” seasonal bins
  ◦ Generalised Pareto model for peaks over threshold
  ◦ Model parameters vary smoothly between bins, optimal
    roughness found using cross-validation
  ◦ Multiple extreme value thresholds with non-exceedance
    probabilities between 0.7 and 0.9 considered
  ◦ Bootstrapping for uncertainties
  ◦ Uncertainty in marginal model not propagated
  ◦ Independent marginal models for pair of variable (StlWnd,
    HndWnd, HndWav) and location (0,1,...,13)
       Jonathan                    MSCE                               September 2021   12 / 23
Data    Marginal transformation to Laplace scale

Scatter plots on Laplace scale

             Registered data on Laplace scale: StlWnd (green), HndWnd (orange), HndWav(blue)

      Jonathan                                   MSCE                                September 2021   13 / 23
Method   Conditional extremes, CE

Conditional extremes

                      Y |{ X = x} = α x + xβ Z

 ◦   Asymptotically-motivated, Heffernan and Tawn [2004]
 ◦   X ∼ Lpl, Y ∼ Lpl, and x > u
 ◦   α ∈ [−1, 1], β ∈ (−∞, 1]
 ◦   Z is a residual random variable characterised empirically, or
     estimated assuming Z ∼ N (µ , σ 2 ), so
                      E[Y |{ X = x}] = α x + µ xβ
                      var[Y |{ X = x}] = σ 2 x2β

 ◦ Identifiability of α and µ when β ≈ 1
 ◦ Model fitting means estimating α , β, µ and σ
        Jonathan                   MSCE                           September 2021   14 / 23
Method   Spatial conditional extremes, SCE

Spatial conditional extremes

                        X |{ X0 = x} = α x + xβ Z

  ◦   Shooter et al. [2019], Wadsworth and Tawn [2019]
  ◦   X = ( X1 , X2 , ..., Xq ), are now observed at q points in space
  ◦   All marginal Xk ∼ Lpl, and x > u
  ◦   α j ∈ [−1, 1], β j ∈ (−∞, 1], j = 1, ..., q

                          Z ∼ DL(µ , σ 2 , δ ; Σ)
  ◦ Delta-Laplace (DL) parameters µ j , σ j > 0, δ j > 0, j = 1, ..., q
  ◦ Σ is a (conditional) correlation matrix with powered-exponential
    decay with distance between the q points (with parameters ρ, κ )

  ◦ Model fitting means estimating α , β, µ , σ , δ and ρ, κ
  ◦ α , β, µ , σ , δ vary smoothly with distance

         Jonathan                     MSCE                               September 2021   15 / 23
Method   Multivariate spatial conditional extremes, MSCE

Multivariate spatial conditional extremes
                         X |{ X01 = x} = α x + xβ Z

  ◦ X = ( X11 , X21 , ..., Xq1 , X12 , X22 , ..., Xq2 , ..., X1m , X2m , ..., Xqm ), for m
    quantities observed at q points in space
  ◦ All marginal Xkℓ ∼ Lpl, and x > u
  ◦ α jℓ ∈ [−1, 1], β jℓ ∈ (−∞, 1], j = 1, ..., q, ℓ = 1, 2, ..., m

                             Z ∼ DL(µ , σ 2 , δ ; Σ)
  ◦ Delta-Laplace (DL) residual parameters µ jℓ , σ jℓ > 0, δ jℓ > 0
  ◦ Σ is a (conditional) correlation matrix with powered-exponential
    decay with distance between the q points for m quantities, with
    appropriate cross-decay (with parameters ρ, κ , λ)

  ◦ Model fitting means estimating α , β, µ , σ , δ and ρ, κ , λ
  ◦ α , β, µ , σ , δ vary smoothly with distance for each quantity
         Jonathan                         MSCE                               September 2021     16 / 23
Method   Inference

Inference

 ◦ Adaptive MCMC, Roberts and Rosenthal [2009]
 ◦ Piecewise linear forms for all parameters with distance using nNod
   spatial nodes
 ◦ Total of m(5nNod + (3m + 1)/2) parameters
 ◦ Rapid convergence, 10k iterations sufficient

      Jonathan                   MSCE               September 2021   17 / 23
Results for North Atlantic     Parameter estimates

Parameter estimates for North Atlantic application

   Estimates for α , β, µ , σ and δ with distance, and residual process estimates ρ, κ and λ. Model fitted with τ = 0.75
                                     StlWnd (green), HndWnd (orange), HndWav(blue)

         Jonathan                                         MSCE                                   September 2021            18 / 23
Results for North Atlantic     Parameter estimates

Laplace-scale conditional mean, standard deviation

  Quantile with threshold probability τ = 0.95 used for illustration. Quantile level is 2.3 at zero distance on green curve
                                   StlWnd (green), HndWnd (orange), HndWav(blue)

          Jonathan                                         MSCE                                   September 2021              19 / 23
Results for North Atlantic   Validation

Laplace-scale simulation under fitted model

                             Threshold probability τ = 0.75
      Jonathan                          MSCE                  September 2021   20 / 23
Results for North Atlantic    Validation

Residual sample

                                                                     ◦ Black = From actual
                                                                       sample
                                                                     ◦ Red = Simulated from
                                                                       fitted model

       Scatter matrix, threshold probability τ = 0.75
     Jonathan                                      MSCE                   September 2021   21 / 23
Summary of findings

Summary of findings
Data
 ◦ JASON and METOP jointly too sparse to use together
 ◦ 1500-2000 good instantaneous daily observations for METOP
 ◦ Sampling bias; swath time roughly the same each day
 ◦ Data are “instantaneous” not storm peak
Methodology
 ◦ Inference straightforward for m = 3 and nNod = 10
 ◦ Assumed distance-dependence structure adopted, particularly for
    residual, seems reasonable from diagnostics
 ◦ Marginal model (fitted independently, uncertainties not pushed
    through to Laplace scale); should do this jointly
Results
 ◦ Results for threshold quantile τ = 0.75, other values examined
 ◦ Conditioning on other locations and quantities examined
 ◦ Spatial extent of extremal dependence for all quantities is about
    600-800km
       Jonathan                        MSCE         September 2021   22 / 23
References

References
O Breivik, O J Aarnes, J-R Bidlot, A Carrasco, and Oyvind Saetra. Wave extremes in the North East Atlantic from ensemble
   forecasts. J. Climate, 26:7525–7540, 2013.
M. G. Genton, S. A. Padoan, and H. Sang. Multivariate max-stable spatial processes. Biometrika, 102:215–230, 2015.
J. E. Heffernan and S. I. Resnick. Limit laws for random vectors with an extreme component. Ann. Appl. Probab., 17:537–571, 2007.
J. E. Heffernan and J. A. Tawn. A conditional approach for multivariate extreme values. J. R. Statist. Soc. B, 66:497–546, 2004.
R. Huser and J. L. Wadsworth. Advances in statistical modelling of spatial extremes. Wiley Interdisciplinary Reviews: Computational
    Statistics, 2020. doi: 10.1002/wics.1537.
Brian J. Reich and Benjamin A. Shaby. A hierarchical max-stable spatial model for extreme precipitation. Ann. Appl. Stat., 6:
    1430–1451, 2012.
A. Ribal and I. R. Young. 33 years of globally calibrated wave height and wind speed data based on altimeter observations. Sci.
    Data, 6:77, 2019.
A. Ribal and I. R. Young. Global calibration and error estimation of altimeter, scatterometer, and radiometer wind speed using
    triple collocation. Remote Sens., 12:1997, 2020.
G. O. Roberts and J. S. Rosenthal. Examples of adaptive MCMC. J. Comp. Graph. Stat., 18:349–367, 2009.
R. Shooter, E. Ross, J. A. Tawn, and P. Jonathan. On spatial conditional extremes for ocean storm severity. Environmetrics, 30:
    e2562, 2019.
R. Shooter, E Ross, A. Ribal, I. R. Young, and P. Jonathan. Spatial conditional extremes for significant wave height from satellite
    altimetry. Environmetrics, 32:e2674, 2021a.
R Shooter, J A Tawn, E Ross, and P Jonathan. Basin-wide spatial conditional extremes for severe ocean storms. Extremes, 24:
    241–265, 2021b.
S. Vettori. Models and Inference for Multivariate Spatial Extremes. KAUST Research Repository., 2017. URL
    https://doi.org/10.25781/KAUST-G2PH8.
S. Vettori, R. Huser, and M. G. Genton. Bayesian modeling of air pollution extremes using nested multivariate max-stable
    processes. Biometrics, 75:831–841, 2019.
J. L. Wadsworth and J. A. Tawn. Higher-dimensional spatial extremes via single-site conditioning. In submission,
      arxiv.org/abs/1912.06560, 2019.
              Jonathan                                         MSCE                                   September 2021               23 / 23
You can also read